FLUX from Black Forest Labs has become the go-to model for AI image generation, surpassing Stable Diffusion in quality and prompt adherence. But it’s also more VRAM-hungry. This guide covers the exact VRAM requirements for every FLUX variant, with real benchmarks and optimization tips to run it on GPUs from 8GB to 32GB.
Quick Navigation:
FLUX Models Overview
Background: FLUX was created by Black Forest Labs, founded by the original creators of Stable Diffusion. FLUX uses a 12B parameter Diffusion Transformer (DiT) architecture, which produces higher-quality images than SD but requires more VRAM.
Schnell
Speed-optimized. 4 inference steps. Open source (Apache 2.0).
Dev
Quality-focused. 20-50 steps. Non-commercial license.
Pro
Production-grade. API only. Commercial use.
VRAM Requirements
FLUX consists of two main components: the text encoder (T5-XXL, ~4.5B params) and the DiT model (~12B params). Together at FP16 they need ~33GB VRAM for maximum speed. Quantization can cut this dramatically.
FLUX Dev/Schnell VRAM by Precision
| Precision | VRAM Used | Quality | Min GPU |
|---|---|---|---|
| FP16 (full quality) | ~33 GB | Best | RTX 5090 32GB |
| FP8 | ~17 GB | Near-perfect | RTX 4090 24GB |
| Q8 (8-bit quantized) | ~13 GB | Excellent | RTX 4070 Ti 16GB |
| NF4 (4-bit quantized) | ~8-10 GB | Good | RTX 3060 12GB |
| Q4 + FP8 T5 | ~6-8 GB | Acceptable | RTX 3060 8GB* |
Pro tip: FP8 is the sweet spot. You get nearly identical image quality to FP16 while using almost half the VRAM. If you have a 24GB GPU, FP8 is the way to go.
VRAM by Resolution
| Resolution | FP16 | FP8 | NF4 |
|---|---|---|---|
| 512x512 | ~28 GB | ~14 GB | ~7 GB |
| 1024x1024 | ~33 GB | ~17 GB | ~10 GB |
| 1536x1536 | ~40 GB+ | ~22 GB | ~13 GB |
GPU Benchmarks
Generation Speed at 1024x1024
Tested with ComfyUI, 20 steps for Dev, 4 steps for Schnell.
| GPU | VRAM | FLUX Dev (FP8) | FLUX Schnell | Verdict |
|---|---|---|---|---|
| RTX 5090 | 32 GB | ~8 sec | ~2 sec | Fastest |
| RTX 4090 | 24 GB | ~14 sec | ~5 sec | Excellent |
| RTX 5080 | 16 GB | ~18 sec (Q8) | ~6 sec | Good |
| RTX 5070 Ti | 16 GB | ~22 sec (Q8) | ~8 sec | Solid |
| RTX 3060 | 12 GB | ~60+ sec (NF4) | ~20 sec | Slow but works |
Running FLUX on Low VRAM
Don’t have a 24GB GPU? You can still run FLUX. Here are the key optimizations:
Use NF4 Quantized Models
NF4 quantization reduces FLUX to ~8-10GB VRAM with surprisingly good quality. Available through ComfyUI and SD Forge WebUI.
Model Offloading + Aggressive Quantization
Use CPU offloading for the text encoder while keeping the DiT on GPU. Combined with Q4 quantization, FLUX can run on GPUs as low as 6GB, though it will be slow (2+ minutes per image).
Use FLUX Schnell Instead of Dev
FLUX Schnell uses only 4 denoising steps vs Dev’s 20-50. This means significantly less VRAM overhead during generation and 5-10x faster output. Quality is slightly lower but excellent for iteration.
GPU Recommendation Summary
Best Overall for FLUX
RTX 4090 (24 GB) or RTX 5090 (32 GB)
Run FP8 FLUX at full speed with room for LoRA training. The 4090 is now the value king on the used market; the 5090 is for those wanting FP16 and future headroom.
Best Budget Option
Runs FLUX Q8 comfortably at 1024x1024. Good balance of price and performance for hobbyists and artists who generate images regularly.
Entry Level
The minimum for a reasonable FLUX experience with NF4 quantization. Slow but functional. Available for under $250 used.
Ready to Generate?
Building a dedicated image gen rig? Our AI Workstation Guide covers the full build process.